| 王晓华,谭石,马大庆.人工神经网络对孤立性肺结节的诊断价值[J].中国医学影像技术,2008,24(7):1114~1117 |
| 人工神经网络对孤立性肺结节的诊断价值 |
| Diagnostic value of artificial neural network in solitary pulmonary nodules |
| 投稿时间：2008-02-03 修订日期：2008-05-27 |
| DOI： |
| 中文关键词: 硬币病变,肺 神经网络 体层摄影术,X线计算机 诊断,鉴别 |
| 英文关键词:Coin lesion, pulmonary Neural networks Tomography, X-ray computed Diagnosis, differential |
| 基金项目:北京市自然科学基金资助项目(7062020)。 |
|
| 摘要点击次数: 2371 |
| 全文下载次数: 680 |
| 中文摘要: |
| 目的 利用人工神经网络(ANN)理论,建立一种全新的模式判别方法用于薄层CT上原发良恶性孤立性肺结节(SPN)的鉴别,并探讨其诊断价值及对不同级别医师的辅助诊断作用。方法 收集经手术或穿刺活检病理证实的SPN 200例(周围型小肺癌 135例,良性结节65例),观察3项临床指标和9项薄层CT指标,并对定性指标进行量化。从中随机选择70%左右的样本(140例)作为训练集,建立ANN诊断模型,并与社会科学统计软件(SPSS)分析处理的Logistic回归模型作比较,计算两种模型对所有样本诊断的正确率和ROC曲线下面积。利用训练好的ANN模型对另外60例样本进行测试,分析初、中、高三级放射科医师使用ANN模型前后的ROC曲线下面积。结果 ANN模型诊断的正确率为98.0%,高于Logistic回归模型的正确率86.0%(P<0.001);两种模型诊断的ROC曲线下面积分别为0.996±0.004和0.936±0.017,差异有统计学意义(P<0.001)。ANN模型及初、中、高级医师首次诊断的ROC曲线下面积分别为0.954、0.737、0.813、0.874,其中ANN与初、中级医师的差异具有统计学意义(P值分别为0.001、0.007),而与高级医师的差异无显著性(P=0.070);初、中、高级医师使用ANN后的ROC曲线下面积分别为0.920、0.938、0.952,三级医师使用ANN后的诊断能力均有显著提高(P值分别为0.000、0.001、0.039);使用ANN后各级医师对SPN的诊断差异无显著性(初-中级、初-高级、中-高级比较的P值分别为0.614、0.369、0.645)。结论 ①根据本研究提出的SPN的征象分类可用于建立ANN模型;②ANN模型用于薄层CT上原发良恶性SPN的鉴别诊断优于传统的Logistic回归模型;③ANN模型对于不同级别的放射科医师都有一定程度的辅助诊断作用。 |
| 英文摘要: |
| Objective To establish a new-type method in differentiating benign from malignant solitary pulmonary nodule (SPN) on thin-slice CT using artificial neural networks (ANN) theory, and to evaluate its diagnostic value and the aided role to different level radiologists. Methods Two hundred cases with pathologically proved SPN by operation or biopsy (small peripheral lung cancer 135, benign nodules 65) were collected; 3 clinical characteristics and 9 thin-slice CT characteristics were observed and quantified the qualitative characteristics. About 70% of all cases (140 cases) were selected randomly to form training samples, on which ANN model were built and compared with Logistic regression obtained by SPSS. The diagnostic consistent rates and areas under ROC of the two models were then calculated. The trained ANN model was used to test the other 60 cases, and the areas under ROC before and after ANN were analyzed in three different level radiologists. Results The total consistent rate of ANN was greater than that of Logistic model (98.0% vs 86.0%, P<0.001). Areas under ROC curve were 0.996±0.004 and 0.936±0.017, respectively, and the difference between the two models was significant (P<0.001). The areas under ROC curve in ANN model, the junior, middle and senior radiologists without ANN were 0.954, 0.737, 0.813 and 0.874, respectively, and the difference between ANN model and the junior, middle radiologists were significant (P=0.001, P=0.007, respectively), while the difference between ANN model and the senior radiologists was not significant (P=0.070). The areas under ROC curve in the junior, middle and senior radiologists with ANN were 0.920, 0.938, and 0.952, respectively, and the performance of junior, middle and senior radiologists with ANN were significantly improved (P<0.001, P=0.001, P=0.039, respectively). The difference of ability to diagnose SPN among different level radiologists with ANN was not significant (P=0.614 for junior-middle, P=0.369 for junior-senior, P=0.645 for middle-senior). Conclusion ①The types of elementary CT signs of SPN in the study can be used to establish ANN model. ②ANN model is superior to traditional logistic model in differentiating primary benign from malignant SPN on thin-slice CT. ③ANN model plays an effective role in aiding different level radiologists to make decisions. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
